Robin Li revealed that the Wenxin Large Model 3.5 version increased the training speed by 2 times and the reasoning speed by 17 times

On June 26, Robin Li, founder, chairman and CEO of Baidu, attended the "Nishan Dialogue on Digital Civilization at the World Internet Conference" and delivered a speech entitled "Large Models Reshape the Digital World".

Large-scale models are the focus of global technological innovation and the main battlefield of global artificial intelligence competition. Li Yanhong believes that "the key point of the new international competition strategy is not how many large models a country has, but how many native AI applications on your large models, and the extent to which these applications have improved production efficiency. If we can squeeze Go to the poker table and get tickets to the competition, China will have a stronger digital industry, and the scale of the digital economy will grow tremendously."

Li Yanhong also revealed in his speech that the Baidu Wenxin model has been iterated to version 3.5. Compared with version 3.0, the training speed has been increased by 2 times, the inference speed has been increased by 17 times, and the model effect has been improved by more than 50%. "Version 3.5 of the Wenxin Large Model is not only a technical upgrade, but also a security upgrade." Li Yanhong emphasized, "The data quality, generation effect and content security have all been significantly improved."

Li Yanhong believes that the deep integration of the digital economy driven by large models and the real economy will make the real economy stronger, better and bigger. In industries such as automobile manufacturing, energy, and transportation, large models can go deep into core business scenarios, and innovate in areas such as intelligent customer service, supply chain, and system scheduling, and promote the digital transformation and intelligent improvement of the industry.

The governance challenges posed by large AI models cannot be ignored. Li Yanhong said in this regard, "Only by insisting on technological development and safe and controllable two-wheel drive can we make steady progress. If we steer the road of AI development safely and responsibly, big models will reshape the digital world, and artificial intelligence can provide China's economy, and even the global economy, create unparalleled prosperity and improve the well-being of all mankind."

The following is the transcript of the speech:

Distinguished leaders and distinguished guests, good morning!

It is a great pleasure to participate in the Nishan Dialogue on Digital Civilization at the World Internet Conference. The topic of my speech is "Large Models Reshape the Digital World".

In the past year, artificial intelligence has advanced at an iterative speed of "weeks" at all levels of technology, products, and applications. The large model has successfully compressed human cognition of the world, allowing us to see the path to general artificial intelligence. The next frontier of large-scale model development is not only to imitate human beings and complete the "prescribed actions" of human beings, but also to help human beings to research and discover unknown areas and break through the limits that human beings have not broken through in the past. If you can take this step, it will be even more meaningful.

How can large models reshape the digital world? I want to talk about it from two levels of technology and application:

At the technical level, in the era of artificial intelligence, the IT technology stack has undergone fundamental changes, from the original three-tier architecture of chips, operating systems, and applications to a four-tier architecture of chips, frameworks, models, and applications:

The bottom layer is the chip layer, and the mainstream chip has changed from CPU to GPU. On top of the chip is the framework layer. The mainstream frameworks include Baidu Fei Paddle, Meta's PyTorch, and Google's TensorFlow. Above the framework is the model layer, and ChatGPT and Wenxin large model are on the model layer. The large model has become the operating system in the era of artificial intelligence, and all applications will be developed based on the large model. Above the model is the application layer, including various AI native applications.

Structural changes in the IT technology stack mean that artificial intelligence, especially large-scale model technology, will restructure the global digital industry. The key point of the new international competition strategy is not how many large models a country has, but how many native AI applications are on your large models, and the extent to which these applications have improved production efficiency. If we can squeeze into the poker table and get tickets to the competition, China will have a stronger digital industry, and the scale of the digital economy will grow tremendously.

Baidu has been investing in artificial intelligence for more than 10 years. It has a full-stack layout in the four layers of chips, frameworks, models, and applications. In terms of key core technologies, Baidu has self-developed leading products and technologies in the four-layer architecture, so it can Perform end-to-end optimization to rapidly improve the efficiency of large model training and inference. The Wenxin large model is completely autonomous and controllable. We have achieved controllable data, controllable framework, and controllable models.

Of course, the governance challenges brought about by large AI models cannot be ignored. The application of new technologies often precedes norms, and only by establishing and improving laws and regulations, institutional systems, and ethics to ensure the healthy development of artificial intelligence can a good innovation ecology be created. Focusing on the future, while paying attention to risk prevention, we should also establish error tolerance and error correction mechanisms at the same time, and strive to achieve a dynamic balance between regulation and development.

Large models are all the rage right now. But 4 years ago, when the large model had not received widespread attention, Baidu launched the Wenxin Large Model 1.0. Then continue to evolve to versions 2.0 and 3.0.

Today, the Wenxin large model has been iterated to version 3.5. Compared with version 3.0 in March, the training speed has increased by 2 times, the inference speed has increased by 17 times, and the cumulative effect of the model has increased by more than 50%.

Wenxin Large Model Version 3.5 is not only a technical upgrade, but also a safety upgrade. We use the industry's mainstream large-scale model basic capability assessment method to carry out the evaluation. The results show that Wenxin large-scale model version 3.5 has been significantly improved in terms of data quality, generation effect, and content security.

my country's artificial intelligence model has a certain foundation, and we need to catch up. At the same time, we should give full play to the advantages of application scenarios, further develop vertical fields, create professional large-scale models in the fields of finance, medical care, and electric power, realize technical optimization with high-quality applications and data feedback, help large-scale models iteratively upgrade, and build an AI ecosystem.

It is foreseeable that large-scale models will penetrate into more and more fields. The digital economy driven by large-scale models will be deeply integrated with the real economy, making the real economy stronger, better and bigger, creating considerable incremental value, bringing In the coming economic and social development and profound changes in the industry.

For example, in the automobile manufacturing industry, the most complex design process requires experienced engineers to find various combinations that meet the needs among more than 20,000 parts and hundreds of thousands of parameters, and then write documents and draw drawings. In Changan Automobile, the large model can efficiently find combination information, automatically generate design documents, and greatly reduce the development cycle and cost. In Sinopec and China Southern Power Grid, large models can go deep into core business scenarios, and innovate in areas such as intelligent customer service, supply chain, and system scheduling, and promote the digital transformation and intelligent improvement of the industry.

In the field of transportation, the intelligent transportation scheme supported by large-scale model technology can improve the efficiency of traffic operation.

For example, on the last working day before the May Day holiday this year, the urban congestion index in Beijing increased by 2.5 times. From the second ring road to the sixth ring road, it is red, and the only green one is Yizhuang. The traffic flow in Yizhuang has also increased significantly, but because of the deployment of the AI global information control solution, more than 300 intelligent intersections can automatically adjust the traffic lights according to the traffic flow, and Yizhuang has become an "oasis" without traffic jams. On the day before the Dragon Boat Festival, the traffic in Beijing urban area and Yizhuang is surprisingly similar: the urban area is very congested, but Yizhuang is smooth.

In the Mount Tai scenic spot in Tai'an, Shandong, in order to serve the development of the tourism economy, promote urban congestion and smooth traffic, and solve the pain point of "difficult parking" for foreign tourists, Baidu used intelligent control methods such as traffic guidance screens and green wave belts to effectively protect citizens and tourists safe travel.

Baidu intelligent transportation solutions have been adopted by 69 cities. By intelligently adjusting the time of traffic lights, traffic efficiency can be increased by 15%-30%, which will drive GDP growth by 2.4%-4.8%.

In Jinan, Shandong, we also landed Baidu Smart Cloud (Shandong) Artificial Intelligence Basic Data Industry Base, which not only cultivated new professions such as AI trainers, but also incubated 22 data labeling technology companies, driving regional employment while stimulating economic growth.

No matter from the perspective of technological trends or industrial applications, large-scale models are by no means a flash in the pan, but a major technological change that affects human development, an engine that drives global economic growth, and a major strategic opportunity that must not be missed.

Adhere to technological development and safe and controllable two-wheel drive in order to achieve steady and long-term progress. If we steer the path of AI development safely and responsibly, the big model will reshape the digital world, and artificial intelligence can create unparalleled prosperity for the Chinese economy, and even the global economy, and improve the well-being of all mankind.

That's all for my speech, thank you!